Generative Adversarial Networks: Face Generation

Author: Jubin Soni

In this project I have used generative adversarial networks to generate new images of faces.

Dataset

I have used be using two datasets in this project:

Since the celebA dataset is complex and I have done GANs in a project for the first time, I wanted to test your neural network on MNIST before used the CelebA dataset. Running the GANs on MNIST allowed me to see how well my model trains and how much time it took to train.

Another option is using FloydHub, set data_dir to "/input" and use the FloydHub data ID "R5KrjnANiKVhLWAkpXhNBe".

In [1]:
data_dir = './data'

# FloydHub - Use with data ID "R5KrjnANiKVhLWAkpXhNBe"
#data_dir = '/input'


"""
UDACITY
DON'T MODIFY ANYTHING IN THIS CELL
"""
import helper

helper.download_extract('mnist', data_dir)
helper.download_extract('celeba', data_dir)
Found mnist Data
Found celeba Data

Explore the Data

MNIST

As you're aware, the MNIST dataset contains images of handwritten digits. You can view the first number of examples by changing show_n_images.

In [3]:
show_n_images = 25

"""
UDACITY
DON'T MODIFY ANYTHING IN THIS CELL
"""
%matplotlib inline
import os
from glob import glob
from matplotlib import pyplot as plt

mnist_images = helper.get_batch(glob(os.path.join(data_dir, 'mnist/*.jpg'))[:show_n_images], 28, 28, 'L')
plt.imshow(helper.images_square_grid(mnist_images, 'L'), cmap='gray')
Out[3]:
<matplotlib.image.AxesImage at 0x7f5d18a4ff98>

CelebA

The CelebFaces Attributes Dataset (CelebA) dataset contains over 200,000 celebrity images with annotations. Since I am going to be generating faces, without the annotations. Viewing the first number of examples by changing show_n_images.

In [4]:
show_n_images = 25

"""
UDACITY
DON'T MODIFY ANYTHING IN THIS CELL
"""
mnist_images = helper.get_batch(glob(os.path.join(data_dir, 'img_align_celeba/*.jpg'))[:show_n_images], 64, 64, 'RGB')
plt.imshow(helper.images_square_grid(mnist_images, 'RGB'))
Out[4]:
<matplotlib.image.AxesImage at 0x7f5d189c7438>

Preprocessing the Data

Since the project's main focus is on building the GANs, first we'll preprocess the data. The values of the MNIST and CelebA dataset will be in the range of -0.5 to 0.5 of 28x28 dimensional images. The CelebA images has been cropped to remove parts of the image that don't include a face, then resized down to 28x28.

The MNIST images are black and white images with a single color channel while the CelebA images have 3 color channels (RGB color channel).

Building the Neural Network

The necessary building components to build a GAN has been done by implementing the following functions below:

  • model_inputs
  • discriminator
  • generator
  • model_loss
  • model_opt
  • train

Checking the Version of TensorFlow and Access to GPU

This will check to make sure you have the correct version of TensorFlow and access to a GPU

In [5]:
"""
UDACITY
DON'T MODIFY ANYTHING IN THIS CELL
"""
from distutils.version import LooseVersion
import warnings
import tensorflow as tf

# Check TensorFlow Version
assert LooseVersion(tf.__version__) >= LooseVersion('1.0'), 'Please use TensorFlow version 1.0 or newer.  You are using {}'.format(tf.__version__)
print('TensorFlow Version: {}'.format(tf.__version__))

# Check for a GPU
if not tf.test.gpu_device_name():
    warnings.warn('No GPU found. Please use a GPU to train your neural network.')
else:
    print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))
TensorFlow Version: 1.1.0
Default GPU Device: /gpu:0

Input

Implementing the model_inputs function to create TF Placeholders for the Neural Network. It should create the following placeholders:

  • Real input images placeholder with rank 4 using image_width, image_height, and image_channels.
  • Z input placeholder with rank 2 using z_dim.
  • Learning rate placeholder with rank 0.

Returning the placeholders in the following the tuple (tensor of real input images, tensor of z data)

In [6]:
import problem_unittests as tests

def model_inputs(image_width, image_height, image_channels, z_dim):
    """
    Create the model inputs
    :param image_width: The input image width
    :param image_height: The input image height
    :param image_channels: The number of image channels
    :param z_dim: The dimension of Z
    :return: Tuple of (tensor of real input images, tensor of z data, learning rate)
    """
    # TODO: Implement Function
    inputs_real = tf.placeholder(tf.float32, shape=(None, image_width, image_height, image_channels), name='input_real') 
    inputs_z = tf.placeholder(tf.float32, (None, z_dim), name='input_z')
    
    learning_rate = tf.placeholder(tf.float32, name='learning_rate')
    
    return inputs_real, inputs_z, learning_rate

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_model_inputs(model_inputs)
Tests Passed

Discriminator

Implementing discriminator to create a discriminator neural network that discriminates on images. This function should be able to reuse the variables in the neural network. Use tf.variable_scope with a scope name of "discriminator" to allow the variables to be reused. The function should return a tuple of (tensor output of the discriminator, tensor logits of the discriminator).

In [7]:
def discriminator(images, reuse=False):
    """
    Create the discriminator network
    :param images: Tensor of input image(s)
    :param reuse: Boolean if the weights should be reused
    :return: Tuple of (tensor output of the discriminator, tensor logits of the discriminator)
    """
    # TODO: Implement Function
    alpha = 0.2
    
    with tf.variable_scope('discriminator', reuse=reuse):
        # using 4 layer network as in DCGAN Paper
        
        # Conv 1
        conv1 = tf.layers.conv2d(images, 64, 5, 2, 'SAME')
        lrelu1 = tf.maximum(alpha * conv1, conv1)
        
        # Conv 2
        conv2 = tf.layers.conv2d(lrelu1, 128, 5, 2, 'SAME')
        batch_norm2 = tf.layers.batch_normalization(conv2, training=True)
        lrelu2 = tf.maximum(alpha * batch_norm2, batch_norm2)
        
        # Conv 3
        conv3 = tf.layers.conv2d(lrelu2, 256, 5, 1, 'SAME')
        batch_norm3 = tf.layers.batch_normalization(conv3, training=True)
        lrelu3 = tf.maximum(alpha * batch_norm3, batch_norm3)
        
        # Conv 4
        conv4 = tf.layers.conv2d(lrelu3, 512, 5, 1, 'SAME')
        batch_norm4 = tf.layers.batch_normalization(conv4, training=True)
        lrelu4 = tf.maximum(alpha * batch_norm4, batch_norm4)
       
        # Flatten
        flat = tf.reshape(lrelu4, (-1, 7*7*512))
        
        # Logits
        logits = tf.layers.dense(flat, 1)
        
        # Output
        out = tf.sigmoid(logits)
        
        return out, logits



"""
UDACITY
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_discriminator(discriminator, tf)
Tests Passed

Generator

Implementing generator to generate an image using z. This function should be able to reuse the variables in the neural network. Use tf.variable_scope with a scope name of "generator" to allow the variables to be reused. The function should return the generated 28 x 28 x out_channel_dim images.

In [8]:
def generator(z, out_channel_dim, is_train=True):
    """
    Creating the generator network
    :param z: Input z
    :param out_channel_dim: The number of channels in the output image
    :param is_train: Boolean if generator is being used for training
    :return: The tensor output of the generator
    """
    # TODO: Implement Function
    alpha = 0.2
    
    with tf.variable_scope('generator', reuse=False if is_train==True else True):
        # Fully connected
        fc1 = tf.layers.dense(z, 7*7*512)
        fc1 = tf.reshape(fc1, (-1, 7, 7, 512))
        fc1 = tf.maximum(alpha*fc1, fc1)
        
        # Starting Conv Transpose Stack
        deconv2 = tf.layers.conv2d_transpose(fc1, 256, 3, 1, 'SAME')
        batch_norm2 = tf.layers.batch_normalization(deconv2, training=is_train)
        lrelu2 = tf.maximum(alpha * batch_norm2, batch_norm2)
        
        deconv3 = tf.layers.conv2d_transpose(lrelu2, 128, 3, 1, 'SAME')
        batch_norm3 = tf.layers.batch_normalization(deconv3, training=is_train)
        lrelu3 = tf.maximum(alpha * batch_norm3, batch_norm3)
        
        deconv4 = tf.layers.conv2d_transpose(lrelu3, 64, 3, 2, 'SAME')
        batch_norm4 = tf.layers.batch_normalization(deconv4, training=is_train)
        lrelu4 = tf.maximum(alpha * batch_norm4, batch_norm4)
        
        # Logits
        logits = tf.layers.conv2d_transpose(lrelu4, out_channel_dim, 3, 2, 'SAME')
        
        # Output
        out = tf.tanh(logits)
        return out


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_generator(generator, tf)
Tests Passed

Loss

Implementing model_loss to build the GANs for training and calculate the loss. The function should return a tuple of (discriminator loss, generator loss). Use the following functions you implemented:

  • discriminator(images, reuse=False)
  • generator(z, out_channel_dim, is_train=True)
In [9]:
def model_loss(input_real, input_z, out_channel_dim):
    """
    Get the loss for the discriminator and generator
    :param input_real: Images from the real dataset
    :param input_z: Z input
    :param out_channel_dim: The number of channels in the output image
    :return: A tuple of (discriminator loss, generator loss)
    """
    # TODO: Implement Function
 
    g_model = generator(input_z, out_channel_dim)
    d_model_real, d_logits_real = discriminator(input_real)
    d_model_fake, d_logits_fake = discriminator(g_model, reuse=True)
    d_loss_real = tf.reduce_mean(
                      tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_real, 
                                                              labels=tf.ones_like(d_model_real) * 0.9))
    d_loss_fake = tf.reduce_mean(
                      tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, 
                                                              labels=tf.zeros_like(d_model_fake)))
    d_loss = d_loss_real + d_loss_fake

    g_loss = tf.reduce_mean(
                 tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake,
                                                         labels=tf.ones_like(d_model_fake)))
    
    return d_loss, g_loss

"""
UDACITY
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_model_loss(model_loss)
Tests Passed

Optimization

Implementing model_opt to create the optimization operations for the GANs. Use tf.trainable_variables to get all the trainable variables. Filter the variables with names that are in the discriminator and generator scope names. The function should return a tuple of (discriminator training operation, generator training operation).

In [10]:
def model_opt(d_loss, g_loss, learning_rate, beta1):
    """
    Get optimization operations
    :param d_loss: Discriminator loss Tensor
    :param g_loss: Generator loss Tensor
    :param learning_rate: Learning Rate Placeholder
    :param beta1: The exponential decay rate for the 1st moment in the optimizer
    :return: A tuple of (discriminator training operation, generator training operation)
    """
    # TODO: Implement Function
    t_vars = tf.trainable_variables()
    d_vars = [var for var in t_vars if var.name.startswith('discriminator')]
    g_vars = [var for var in t_vars if var.name.startswith('generator')]

    # Optimize
    d_train_opt = tf.train.AdamOptimizer(learning_rate, beta1=beta1).minimize(d_loss, var_list=d_vars)
    g_train_opt = tf.train.AdamOptimizer(learning_rate, beta1=beta1).minimize(g_loss, var_list=g_vars)

    return d_train_opt, g_train_opt

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_model_opt(model_opt, tf)
Tests Passed

Neural Network Training

Function to show Output

Creating a function to show the current output of the generator during training. It will help determine how well the GANs is training.

In [11]:
"""
DON'T MODIFY ANYTHING IN THIS CELL
"""
import numpy as np

def show_generator_output(sess, n_images, input_z, out_channel_dim, image_mode):
    """
    Show example output for the generator
    :param sess: TensorFlow session
    :param n_images: Number of Images to display
    :param input_z: Input Z Tensor
    :param out_channel_dim: The number of channels in the output image
    :param image_mode: The mode to use for images ("RGB" or "L")
    """
    cmap = None if image_mode == 'RGB' else 'gray'
    z_dim = input_z.get_shape().as_list()[-1]
    example_z = np.random.uniform(-1, 1, size=[n_images, z_dim])

    samples = sess.run(
        generator(input_z, out_channel_dim, False),
        feed_dict={input_z: example_z})

    images_grid = helper.images_square_grid(samples, image_mode)
    plt.imshow(images_grid, cmap=cmap)
    plt.show()

Train Function

Implementing train function to build and train the GANs. Use the following functions you implemented:

  • model_inputs(image_width, image_height, image_channels, z_dim)
  • model_loss(input_real, input_z, out_channel_dim)
  • model_opt(d_loss, g_loss, learning_rate, beta1)

Use the show_generator_output to show generator output while you train. Running show_generator_output for every batch will drastically increase training time and increase the size of the notebook. It's recommended to print the generator output every 100 batches.

In [12]:
def train(epoch_count, batch_size, z_dim, learning_rate, beta1, get_batches, data_shape, data_image_mode):
    """
    Train the GAN
    :param epoch_count: Number of epochs
    :param batch_size: Batch Size
    :param z_dim: Z dimension
    :param learning_rate: Learning Rate
    :param beta1: The exponential decay rate for the 1st moment in the optimizer
    :param get_batches: Function to get batches
    :param data_shape: Shape of the data
    :param data_image_mode: The image mode to use for images ("RGB" or "L")
    """
    # TODO: Build Model
    tf.reset_default_graph()
    input_real, input_z, _ = model_inputs(data_shape[1], data_shape[2], data_shape[3], z_dim)
    d_loss, g_loss = model_loss(input_real, input_z, data_shape[3])
    d_optimized, g_optimized = model_opt(d_loss, g_loss, learning_rate, beta1)
    steps=0
    
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        for epoch_i in range(epoch_count):
            for batch_images in get_batches(batch_size):
                # TODO: Train Model
                batch_images = batch_images * 2
                steps += 1
                batch_z = np.random.uniform(-1, 1, size=(batch_size, z_dim))
                
                _ = sess.run(d_optimized, feed_dict={input_real:batch_images, input_z:batch_z})
                _ = sess.run(g_optimized, feed_dict={input_z:batch_z})
                
                if steps % 100 == 0:
                    train_loss_d = d_loss.eval({input_z: batch_z, input_real: batch_images})
                    train_loss_g = g_loss.eval({input_z: batch_z})
                    
                    print('Epoch {}/{}: '.format(epoch_i+1, epochs),
                         'Discriminator Loss: {:.4f}...'.format(train_loss_d),
                         'Generator Loss: {:.4f}...'.format(train_loss_g))
                
                _ = show_generator_output(sess, 1, input_z, data_shape[3], data_image_mode)

Testing on MNIST

As mentioned earlier, First I am testing my GANs architecture on MNIST. After 2 epochs, the GANs should be able to generate images that look like handwritten digits. I have made sure the loss of the generator is lower than the loss of the discriminator or close to 0.

It took more than an hour for the GAN to complete the testing and generate images that looked like real-numbers.

In [ ]:
batch_size = 10
z_dim = 100
learning_rate = 0.0005
beta1 = 0.5


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 2

mnist_dataset = helper.Dataset('mnist', glob(os.path.join(data_dir, 'mnist/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, mnist_dataset.get_batches,
          mnist_dataset.shape, mnist_dataset.image_mode)
Epoch 1/2:  Discriminator Loss: 3.3839... Generator Loss: 3.9083...
Epoch 1/2:  Discriminator Loss: 1.2672... Generator Loss: 0.7696...
Epoch 1/2:  Discriminator Loss: 1.3699... Generator Loss: 2.1905...
Epoch 1/2:  Discriminator Loss: 1.8548... Generator Loss: 2.8744...
Epoch 1/2:  Discriminator Loss: 0.5790... Generator Loss: 2.7744...
Epoch 1/2:  Discriminator Loss: 1.2861... Generator Loss: 2.0159...
Epoch 1/2:  Discriminator Loss: 1.1925... Generator Loss: 1.8313...
Epoch 1/2:  Discriminator Loss: 1.0652... Generator Loss: 0.9692...
Epoch 1/2:  Discriminator Loss: 0.7627... Generator Loss: 1.5023...
Epoch 1/2:  Discriminator Loss: 1.6539... Generator Loss: 0.4254...
Epoch 1/2:  Discriminator Loss: 0.8400... Generator Loss: 1.3389...
Epoch 1/2:  Discriminator Loss: 0.9008... Generator Loss: 1.1769...
Epoch 1/2:  Discriminator Loss: 0.9157... Generator Loss: 1.4336...
Epoch 1/2:  Discriminator Loss: 0.9475... Generator Loss: 1.3516...
Epoch 1/2:  Discriminator Loss: 1.0502... Generator Loss: 2.4816...
Epoch 1/2:  Discriminator Loss: 1.1524... Generator Loss: 0.7083...
Epoch 1/2:  Discriminator Loss: 0.7429... Generator Loss: 1.4369...
Epoch 1/2:  Discriminator Loss: 1.7750... Generator Loss: 0.7160...
Epoch 1/2:  Discriminator Loss: 3.9071... Generator Loss: 0.0626...
Epoch 1/2:  Discriminator Loss: 0.7337... Generator Loss: 1.3965...
Epoch 1/2:  Discriminator Loss: 1.0512... Generator Loss: 1.0214...
Epoch 1/2:  Discriminator Loss: 3.3576... Generator Loss: 0.0864...
Epoch 1/2:  Discriminator Loss: 1.4738... Generator Loss: 0.8524...
Epoch 1/2:  Discriminator Loss: 1.1214... Generator Loss: 1.0998...

Testing on CelebA

Finally, Running my GANs on CelebA. It took around 20 minutes on Amazon AWS average GPU to run one epoch. You can run the whole epoch or stop when it starts to generate realistic faces.

In [13]:
batch_size = 64
z_dim = 100
learning_rate = 0.0009
beta1 = 0.5


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode)
Epoch 1/1:  Discriminator Loss: 1.6473... Generator Loss: 1.2843...
Epoch 1/1:  Discriminator Loss: 1.8825... Generator Loss: 0.3290...
Epoch 1/1:  Discriminator Loss: 0.9856... Generator Loss: 1.0353...
Epoch 1/1:  Discriminator Loss: 1.3032... Generator Loss: 1.5449...
Epoch 1/1:  Discriminator Loss: 1.3988... Generator Loss: 1.6494...
Epoch 1/1:  Discriminator Loss: 1.0494... Generator Loss: 0.8597...
Epoch 1/1:  Discriminator Loss: 1.2676... Generator Loss: 0.9485...
Epoch 1/1:  Discriminator Loss: 1.7360... Generator Loss: 0.3794...
Epoch 1/1:  Discriminator Loss: 1.3304... Generator Loss: 1.0281...
Epoch 1/1:  Discriminator Loss: 1.3711... Generator Loss: 0.9168...
Epoch 1/1:  Discriminator Loss: 0.8664... Generator Loss: 1.7641...
Epoch 1/1:  Discriminator Loss: 1.1845... Generator Loss: 1.3057...
---------------------------------------------------------------------------
KeyboardInterrupt                         Traceback (most recent call last)
/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/ops.py in get_controller(self, default)
   3624       self.stack.append(default)
-> 3625       yield default
   3626     finally:

<ipython-input-13-259e9fa8d093> in <module>()
     14     train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
---> 15           celeba_dataset.shape, celeba_dataset.image_mode)

<ipython-input-12-544edd2a93ef> in train(epoch_count, batch_size, z_dim, learning_rate, beta1, get_batches, data_shape, data_image_mode)
     39 
---> 40                 _ = show_generator_output(sess, 1, input_z, data_shape[3], data_image_mode)

<ipython-input-11-ab89ba32f1f2> in show_generator_output(sess, n_images, input_z, out_channel_dim, image_mode)
     20         generator(input_z, out_channel_dim, False),
---> 21         feed_dict={input_z: example_z})
     22 

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
    777       result = self._run(None, fetches, feed_dict, options_ptr,
--> 778                          run_metadata_ptr)
    779       if run_metadata:

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
    981       results = self._do_run(handle, final_targets, final_fetches,
--> 982                              feed_dict_string, options, run_metadata)
    983     else:

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)
   1031       return self._do_call(_run_fn, self._session, feed_dict, fetch_list,
-> 1032                            target_list, options, run_metadata)
   1033     else:

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in _do_call(self, fn, *args)
   1038     try:
-> 1039       return fn(*args)
   1040     except errors.OpError as e:

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in _run_fn(session, feed_dict, fetch_list, target_list, options, run_metadata)
   1020                                  feed_dict, fetch_list, target_list,
-> 1021                                  status, run_metadata)
   1022 

KeyboardInterrupt: 

During handling of the above exception, another exception occurred:

IndexError                                Traceback (most recent call last)
<ipython-input-13-259e9fa8d093> in <module>()
     13 with tf.Graph().as_default():
     14     train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
---> 15           celeba_dataset.shape, celeba_dataset.image_mode)

/usr/lib/python3.5/contextlib.py in __exit__(self, type, value, traceback)
     75                 value = type()
     76             try:
---> 77                 self.gen.throw(type, value, traceback)
     78                 raise RuntimeError("generator didn't stop after throw()")
     79             except StopIteration as exc:

/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/ops.py in get_controller(self, default)
   3626     finally:
   3627       if self._enforce_nesting:
-> 3628         if self.stack[-1] is not default:
   3629           raise AssertionError(
   3630               "Nesting violated for default stack of %s objects"

IndexError: list index out of range

Stopping the Training

I have stopped the Testing of CelebA and MNIST after I felt that the neural network has started generating the images correctly. However, the generating of numbers and faces is taking a lot of time and I am running this job on AWS GPU, and it has already been > 6 hours and to avoid large bill I am stopping the testing early.